seedmanc's picture
Update opp.py
12d2a52 verified
# -*- coding: utf-8 -*-
"""batch aesthetics predictor v2 - release.ipynb
Automatically generated by Colaboratory.
Original file is located at
https://colab.research.google.com/drive/1zTrHop7pStcCwPAUP_nekK1rp6lcppYx
"""
# Commented out IPython magic to ensure Python compatibility.
# %%capture
# #@title Install environment & dl MLP { form-width: "100%", display-mode: "form" }
# !pip install git+https://github.com/openai/CLIP.git
# !pip install gradio~=3.18.0
# #!pip install torch==1.13.1#+cu116
# !pip install pytorch-lightning~=2.0.1
# !wget -nc https://huggingface.co/spaces/Seedmanc/batch-laion-aesthetic-predictor/resolve/main/sac%2Blogos%2Bava1-l14-linearMSE.pth
#@title CLIP dl & init { run: "auto", vertical-output: true, form-width: "25%", display-mode: "form" }
checkpoint = "ViT-L/14" #@param ["ViT-L/14", "ViT-L/14@336px"]
import numpy as np
import torch
import pytorch_lightning as pl
import torch.nn as nn
import clip
import time
global prev_time
global isCpu
# if you changed the MLP architecture during training, change it also here:
class MLP(pl.LightningModule):
def __init__(self, input_size, xcol='emb', ycol='avg_rating'):
super().__init__()
self.input_size = input_size
self.xcol = xcol
self.ycol = ycol
self.layers = nn.Sequential(
nn.Linear(self.input_size, 1024),
#nn.ReLU(),
nn.Dropout(0.2),
nn.Linear(1024, 128),
#nn.ReLU(),
nn.Dropout(0.2),
nn.Linear(128, 64),
#nn.ReLU(),
nn.Dropout(0.1),
nn.Linear(64, 16),
#nn.ReLU(),
nn.Linear(16, 1)
)
def forward(self, x):
return self.layers(x)
def normalized(a, axis=-1, order=2):
l2 = np.atleast_1d(np.linalg.norm(a, order, axis))
l2[l2 == 0] = 1
return a / np.expand_dims(l2, axis)
def load_models():
model = MLP(768)
global device
device = "cuda" if torch.cuda.is_available() else "cpu"
global isCpu
isCpu = device == "cpu"
s = torch.load("sac+logos+ava1-l14-linearMSE.pth", map_location=device)
model.load_state_dict(s)
model.to(device)
model.eval()
model2, preprocess = clip.load(checkpoint, device=device, jit=True)
model_dict = {}
model_dict['classifier'] = model
model_dict['clip_model'] = model2
model_dict['clip_preprocess'] = preprocess
model_dict['device'] = device
return model_dict
if __name__ == '__main__':
print('\tinit models')
global model_dict
prev_time = time.time()
model_dict = load_models()
print('model load', time.time() - prev_time)
description = f"""
## Batch Image Aesthetic Predictor
0. Based on https://huggingface.co/spaces/Geonmo/laion-aesthetic-predictor, I just expanded the GUI & added stats.
1. This model is designed by adding five MLP layers on top of (frozen) CLIP <u>**{checkpoint}**</u> checkpoint and only the MLP layers are fine-tuned with a lot of images by a regression loss term such as MSE and MAE.
2. Output is bounded from 0 to 10. The higher the better.
3. The MLP being used currently is: **sac+logos+ava1-l14-linearMSE.pth** trained on 224x224 images.
4. Running on **{device}**{', be patient. Progressive output & immediate stats are available.' if isCpu else '. Batch mode enabled, results after completion.'}
5. Please don't click 'Submit' again during the processing, it'll mess things up. To stop processing, clear the file input. If the results are missing from the stats or export areas at the end, sort the table by any header & wait.
{'6. The MLP was not retrained for this CLIP checkpoint, correct results are not guaranteed. It is also 2x slower.' if checkpoint != "ViT-L/14" else ''}
"""
#@title 👁️⃤ { run: "auto", form-width: "15%" }
global predict#or
writeClip = False #param {type:"boolean"}
import os
from PIL import Image
if writeClip: #disabled in v1
import torchvision
os.makedirs('CLIPped', exist_ok=True)
def predict(image):
img_input = model_dict['clip_preprocess'](Image.open(image))
clipped = None
if writeClip:
clipped = img_input
image_input = img_input.unsqueeze(0).to(model_dict['device']) #try batch
with torch.no_grad():
image_features = model_dict['clip_model'].encode_image(image_input)
if model_dict['device'] == 'cuda': # add TPU support?
im_emb_arr = normalized(image_features.detach().cpu().numpy())
im_emb = torch.from_numpy(im_emb_arr).to(model_dict['device']).type(torch.cuda.FloatTensor)
else:
im_emb_arr = normalized(image_features.detach().numpy())
im_emb = torch.from_numpy(im_emb_arr).to(model_dict['device']).type(torch.FloatTensor)
prediction = model_dict['classifier'](im_emb)
score = prediction.item() #optimize?
return score, clipped
#@title Wrapper & stats { form-width: "10%" }
DEBUG = True #@param {type:"boolean"}
autoclearLog = True #@param {type:"boolean"}
import csv
import sys
import gradio as gr
if DEBUG: print(gr.__version__) #
def defStats():
return {'Max':{}, 'Max - min': {}}
global Ready
global avgScore
global eta
global speed
global canPoll
canPoll=Ready =False
eta = avgScore = None
speed = 0
Stats = defStats()
global default_mode
default_mode = list(Stats.keys())[1]
def log(x = '', y = None): # debug only
if not DEBUG:
return x
global prev_time
print(f"<\033[97m{sys._getframe().f_back.f_code.co_name}\033[96m>:")
if x:
print(time.strftime('%M:%S '), x, round(time.time() - prev_time, 3), '\033[0m')
if y:
print(' extra: ', y, '\033[0m')
prev_time = time.time()
return x
def pollStatus(table=[]): ### TODO idk what to do
time.sleep(1)
spd = speed and (f'{round(speed,1)} s/f' if speed >= 1 else f'{round(1/speed,1)} f/s')
stext = f' Avg speed: {spd}.' if speed else ''
etext = f' ETA: {eta} {"s." if type(eta) == int else ""}' if eta else ''
atext = f'Running average: {avgScore}.' if avgScore else ''
return f"[{time.strftime('%M:%S')}] {' Ready.' if not atext else ''} {atext} {etext} {stext}" if canPoll else 'idle'
def switch_stats(mode):
global default_mode
default_mode = mode if mode else 'Max'
return Stats[default_mode]
def writeStats(labels):
with open('stats.csv', 'w', newline='') as f:
writer = csv.writer(f)
log('actual write stats', labels and labels.values()) #
writer.writerow(gr.utils.sanitize_list_for_csv(labels.keys()))
writer.writerow(gr.utils.sanitize_list_for_csv(labels.values()))
# MAIN ################################################################
def batch_predict(files=None, progress=gr.Progress()): #=> stats_toggle, stats_output, table_output, submit_btn
run_time = time.time()
if files and len(files) > 1:
global eta
eta = 'calculating...'
results = list()
log('has file(s)?', files and files[0])
global Stats
global Ready
Stats = defStats()
if files is None:
log('empty load')
yield gr.update(), None, None, gr.update()
log('ABORT')
return
else:
maxSteps = min(len(files), 3) if isCpu else len(files)
log('good2go')
yield gr.update(visible=False), gr.update(visible=False), None, gr.update(variant="secondary")
progress((1, maxSteps), unit='', desc='Importing...')
clearStats()
log('start the main loop')
times=list()
clips=list()
for idx,file in enumerate(files, start=1):
prev_time = time.time()
score, clipped = predict(file)
if not Ready: # the solution to the interruption bug, do not remove #
return
results.append([file.orig_name, round(score, 5), None])
if writeClip: #disabled in v1
clips.append((clipped, file.orig_name))
times.append(time.time() - prev_time) #simplify
asyncThreshold = 1 if isCpu else len(files)-1
if (idx <= asyncThreshold):
progress((idx+1, maxSteps), unit='', desc='Starting...' if isCpu else 'Working...')
if (idx > asyncThreshold) and (idx < len(files)): # === False if not isCpu
global avgScore
global speed
speed = np.mean(times)
avgScore = statistics(results, False)
eta = round(speed*(len(files)-idx+1)) # +1 or [1::]?
log(idx)
yield gr.update(), None, results, gr.update()
table_data = results
if DEBUG: print('RUN time', time.time() - run_time, 'avg', np.mean(times)) #
if len(results) > 1:
eta = 'finishing...'
log('finishing')
stats = statistics(results)
for i, row in enumerate(table_data):
table_data[i][2] = round((row[1] - stats['AVG'])**2, 4) # pylint: disable=report-general-type-issues
writeStats(stats)
log('|2|', table_data) #
yield gr.update(visible=True), gr.update(value=switch_stats(default_mode), visible=True), table_data, gr.update(variant="primary")
else:
log('I', table_data) #
yield gr.update(visible=False), gr.update(value=None, visible=False), table_data, gr.update(variant="primary") #
avgScore = None
if writeClip: #supposedly runs async w/o delaying the results? disabled in v1 anyway
log('beforeWrite')
for c,f in clips:
torchvision.utils.save_image(c, 'CLIPped/'+f+'.png', normalize=True)
log('afterWrite')
log('Exit main loop')
speed = (time.time() - run_time)/len(files)
# /main #####################################################################
def statistics(results, full=True):
array = np.array(results).T[1].astype(float)
max = np.max(array)
avg = round(array.mean(), 3)
if (not full): return avg
med = round(np.median(array), 3)
min = array.min()
std = round(array.std(), 4)
cov = round(std/avg*100, 2)
rng = round(max-min, 3)
range = max-min
Stats['Max'][f'MAX: {round(max, 3)}'] = 1
Stats['Max'][f'min: {round(min, 3)}'] = min/max
Stats['Max'][f"CoV: {cov}%"] = std/max
Stats['Max'][f'AVG: {avg}'] = avg/max
Stats['Max'][f'Med: {med}'] = med/max
Stats['Max'][f'M-m: {rng}'] = range/max
# TODO can this be shortened?
if (range == 0):
range = 1
Stats['Max - min'][f'MAX: {round(max, 3)}'] = 1
Stats['Max - min'][f'min: {round(min, 3)}'] = 0
Stats['Max - min'][f"CoV: {cov}%"] = std/range
Stats['Max - min'][f'AVG: {avg}'] = (avg-min)/range
Stats['Max - min'][f'Med: {med}'] = (med-min)/range
Stats['Max - min'][f'M-m: {rng}'] = rng/max
return dict(zip(('AVG','CoV','M-m','min','Med','MAX'), (avg, cov, rng, round(min,3), med, round(max,3))))
def clearStats():
log('clst too many calls?') #
for root, dirs, files in os.walk('.'):
for file in files:
if (file.startswith(('scores','stats'))): # TODO separate folder, names?
os.remove(file)
def scan():
r = ['scores.csv', 'stats.csv']
return [x for x in r if os.path.isfile(x)]
# buggy as fuck
def writeScores(table, files): # => csv_output, stats_output, stats_toggle
statsVisible = False
rows = table and table['data']
log('Entering the scores writer', 'from table change' if files and table else None)
showStats = (gr.update(visible=statsVisible) for x in range(0,2)) # add full return statement?
if files is None:
log('No files, exiting writer')#
resetStatus('from table') # refactor
return [gr.update(value=scan()), *list(showStats)]
######
def writes(tbl):
with open('scores.csv', 'w', newline='') as f: #try tsv, json
writer = csv.writer(f)
log('Actual saving scores', len(tbl['data'])) #
writer.writerow(gr.utils.sanitize_list_for_csv(tbl['headers']))
writer.writerows(gr.utils.sanitize_list_for_csv(tbl['data']))
######
if table and any([x for x in rows[0]]):
if (len(rows) > 1):
statsVisible = len(rows) >= len(files)
if statsVisible:
writes(table)
log('Updating two', 'finished') #
global eta
eta = 0
return [gr.update(value=scan()), *list(showStats)]
else:
statsVisible = False
if (len(files) == 1):
writes(table)
log('updating 1') #
return [gr.update(value=scan()), *list(showStats)]
log('Not ready for writing yet, exiting.', f'total files: {files and len(files)}, but ready rows: {rows and len(rows)}')
return [gr.update(value=scan()), *list(showStats)]
#@title GUI { vertical-output: true, form-width: "50%", display-mode: "both" }
tableQueued_False = False #@param {type:"boolean"}
queueConcurrency_2 = 10 #@param {type:"integer", min:1}
queueUpdateInterval_0 = 0 #@param {type:"slider", min:0, max:10, step:0.2}
#@markdown tableQueued == True + queueConcurrency == 1 guarantees stalling on CPU
#@markdown
#@markdown tableQueued - unknown effect on speed or stability
#@markdown
#@markdown queueConcurrency > 1 - technically should improve speed?
#@markdown
#@markdown queueUpdateInterval - in (0, 1] slows down processing, otherwise seems useless.
#@markdown prevent_thread_lock - keep the "busy cell" behavior of debug mode without it to avoid multiple instances running in parallel;
#@markdown effects on speed & stability unknown
if DEBUG:
import shutil #i doshutilsya
import subprocess
if writeClip: # disabled in v1
for root, dirs, files in os.walk('CLIPped'):
for file in files:
os.remove('CLIPped/'+file)
if DEBUG:
for root, dirs, files in os.walk('../tmp'): #debug only
for dir in dirs:
shutil.rmtree('../tmp/'+dir)
for file in files:
os.remove('../tmp/'+file) #/debug
def resetStatus(msg = 'clear'):
global avgScore
global eta
global speed
avgScore = None
eta = None
speed = 0
log(msg)
if msg != 'clear':
clearStats()
print('\n')
Css = '''
#lbl .output-class {
background-color: transparent;
max-height: 0;
color: transparent;
padding: var(--size-3);
}
#add_img .file-preview .file td:first-child {
overflow-wrap: anywhere;
}
#csv_out .file-preview {
margin-bottom: var(--size-4);
overflow-x: visible;
}
#tbl_out tbody .cell-wrap:first-child {
overflow-wrap: anywhere;
}
button#sbmt:focus:not(:active) {
opacity: 0.75;
pointer-events: none;
}
#mid_col :not(#csv_out) .wrap.default {
opacity: 0!important;
}
'''
def toggleRun(files): # => submit, dataframe, status
global Ready
Ready = files is not None
log('Toggle', Ready)
global canPoll
canPoll = Ready
if not Ready:
if eta:
log('INTERRUPTED at ss remaining (extra)', eta)
resetStatus()
if DEBUG and autoclearLog:
subprocess.call('clear')
print('\r')
clearStats()
return gr.Button.update(variant='primary' if Ready else 'secondary'), None, pollStatus()
# ''', interactive=True''')
log('GUI start')
blks = gr.Blocks(analytics_enabled=False, title="Batch Image Aesthetic Predictor", css=Css)
with blks as demo:
with gr.Accordion('README', open=False):
gr.Markdown(description)
with gr.Row().style(equal_height=False):
with gr.Column(scale=2):
imageinput = gr.Files(file_types=["image"], label="Add images", elem_id="addimg")
submit_button = gr.Button('Submit', variant="secondary", elem_id='sbmt') #TODO interactive
with gr.Column(variant="compact", min_width=256, elem_id="mid_col"):
stats_toggle = gr.Radio(list(Stats.keys()), show_label=True, label='Stats relative to:', value=default_mode, visible=False)
stats_output = gr.Label(label='Stats', visible=False, elem_id="lbl")
csv_output = gr.File( label="Export", elem_id='csv_out' )
with gr.Column(scale=2):
table_output = gr.Dataframe(headers=['Image', 'Score', 'MSE'], max_rows=15, overflow_row_behaviour="paginate", interactive=False, wrap=True, elem_id="tbl_out")
status = gr.Textbox(pollStatus(), max_lines=1, show_label=False, placeholder='Status bar').style(container=False)
status.change(pollStatus, None, status, show_progress= False, queue=False)
tch = table_output.change(writeScores, [table_output, imageinput], [csv_output, stats_output, stats_toggle], preprocess=False, queue= tableQueued_False, show_progress=not isCpu)
stats_toggle.change(switch_stats, [stats_toggle], [stats_output], queue=False, show_progress=False)
run = submit_button.click(batch_predict, imageinput, [stats_toggle, stats_output, table_output, submit_button], queue=True, scroll_to_output=True)
#imageinput.clear(reset, [imageinput], [table_output], queue=False, show_progress=True, preprocess=False)
imageinput.change(toggleRun, imageinput, [submit_button, table_output, status], queue= False, cancels=[run], show_progress=False) #
# try .then()
if DEBUG:
demo.load(lambda: log('load'), queue=not True, show_progress=False)
demo.queue(api_open= not DEBUG, status_update_rate='auto' if queueUpdateInterval_0 == 0 else queueUpdateInterval_0 , concurrency_count=max(queueConcurrency_2, 1))
log('Prelaunch')
demo.launch(debug=DEBUG, quiet=not DEBUG, show_error=True)
#demo.close()